Modeling CAR Response at the Single-Cell Level Using Conditional Optimal Transport

Abstract

Chimeric Antigen Receptor (CAR) T cell therapy is a promising area of cancer immunotherapy. However, many challenges such as loss of persistence, T cell exhaustion, and therapy associated toxicities hamper further advancement of CAR T cell therapy. Therefore, recent efforts have focused on designing improved CARs that show better therapeutic characteristics. However, it is unfeasible to test all CAR variants in lab-based assays as CARs consist of multiple intracellular signalling domains. This results in over 100000 possible variants. We leverage computational modeling to navigate this vast combinatorial space by learning the relationship between CAR design and T cell functionality, thereby proposing promising CAR T cell designs. CAR T cells expressing different variants can be viewed as cells that underwent different perturbations. Neural Optimal Transport is an upcoming field that can model single cell perturbations and predict unseen cells and conditions. In this work we leverage the conditional Monge Gap to model the response to CAR expression at a single-cell level and generate gene expression of cells that express an unseen CAR design. We show that CAR OT (CAROT) significantly outperforms the baseline for gene expression prediction for in-distribution CAR variants, with distinct gene expression patterns per CAR that capture biological characteristics. When predicting unseen CAR variants, we demonstrate promising results in terms of gene expression prediction and show the model learns gene expression patterns linked to domains in the training set. This work demonstrates that optimal transport may support discovery and development of new CAR T cell designs.